[Bug target/59464] New: Unnecessary vector register spill

[Bug target/59464] New: Unnecessary vector register spill

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59464

            Bug ID: 59464
           Summary: Unnecessary vector register spill
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: hjl.tools at gmail dot com
                CC: ubizjak at gmail dot com
            Target: x86-64

[[hjl@gnu-6 gcc]$ cat /tmp/v.c
typedef struct
{
  float f1;
  float f2;
  float f3;
  float f4;
} test_structure_8;
test_structure_8 struct8 (test_structure_8 ts)
{
  ts.f1 += 1;
  ts.f2 += 1;
  ts.f3 += 1;
  ts.f4 += 1;
  return ts;
}
[hjl@gnu-6 gcc]$ ./xgcc -B./ -O3 -S /tmp/v.c
[hjl@gnu-6 gcc]$ cat v.s
    .file    "v.c"
    .section    .text.unlikely,"ax",@progbits
.LCOLDB1:
    .text
.LHOTB1:
    .p2align 4,,15
    .globl    struct8
    .type    struct8, @function
struct8:
.LFB0:
    .cfi_startproc
    movq    %xmm1, -32(%rsp)
    movq    %xmm0, -40(%rsp)
    movups    -40(%rsp), %xmm0
    addps    .LC0(%rip), %xmm0
    movaps    %xmm0, -56(%rsp)
    movq    -48(%rsp), %rax
    movq    -56(%rsp), %xmm0
    movq    %rax, -56(%rsp)
    movq    -56(%rsp), %xmm1
    ret
    .cfi_endproc
.LFE0:
    .size    struct8, .-struct8
    .section    .text.unlikely
.LCOLDE1:
    .text
.LHOTE1:
    .section    .rodata.cst16,"aM",@progbits,16
    .align 16
.LC0:
    .long    1065353216
    .long    1065353216
    .long    1065353216
    .long    1065353216
    .ident    "GCC: (GNU) 4.9.0 20131210 (experimental)"
    .section    .note.GNU-stack,"",@progbits
[hjl@gnu-6 gcc]$ 

At least, we should do

    .globl    struct8
    .type    struct8, @function
struct8:
.LFB0:
    .cfi_startproc
    movq    %xmm1, -32(%rsp)
    movq    %xmm0, -40(%rsp)
    movups    -40(%rsp), %xmm0
    addps    .LC0(%rip), %xmm0
    movaps    %xmm0, -56(%rsp)
    movq    -48(%rsp), %xmm1
    movq    -56(%rsp), %xmm0
    ret
    .cfi_endproc
.LFE0:
    .size    struct8, .-struct8

[Bug target/59464] Unnecessary vector register spill

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59464

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-12-11
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
struct8 (struct test_structure_8 ts)
{
  vector(4) float vect__2.10;
  vector(4) float vect_ts_f1_9.9;
  struct test_structure_8 D.1759;

;;   basic block 2, loop depth 0
;;    pred:       ENTRY
  vect_ts_f1_9.9_15 = MEM[(struct  *)&ts];
  vect__2.10_17 = vect_ts_f1_9.9_15 + { 1.0e+0, 1.0e+0, 1.0e+0, 1.0e+0 };
  MEM[(struct  *)&D.1759] = vect__2.10_17;
  return D.1759;
;;    succ:       EXIT

}

OTOH the code generated is much much worse if you disable vectorization.

Then we should make sure the stack slot we spill the arguments to is
properly aligned so we can avoid the UNSPEC_LOADU.  This eventually
can lead to cleanup opportunities.

Ideally we'd recognize that we can more optimally load from the argument
registers into a V4SF register of course (but due to the way we do argument
setup during expand this may not be easily possible).

It's of course the bad ABI choice that leads to this mess ...